Session Bench v6: New Stuff vs Baseline

Comparison report built from existing benchmark snapshots: v1 baseline vs v5 (latest measured new-stuff snapshot).

Baseline Source

Compared Snapshot

Largest Improvement

JSONL -99.3%

Status

Mixed

Median Latency: Baseline vs New Stuff

Operation	v1 Baseline	v5 New Stuff	Delta	Interpretation
session list	787ms	856ms	+69ms (+8.8%)	slower
cleanup dry-run	383ms	401ms	+18ms (+4.7%)	slower
rg content search	53ms	55ms	+2ms (+3.8%)	near-flat
buffer read (single)	299ms	310ms	+11ms (+3.7%)	slower
tab color cycle	916ms	1004ms	+88ms (+9.6%)	slower
yaml↔iterm cross-ref	310ms	327ms	+17ms (+5.5%)	slower

Key New-Stuff Win (Component Level)

Component	v1 Baseline	v5 New Stuff	Delta	Why
JSONL version lookup (all sessions)	705ms (rglob + parse)	5ms (direct path + tail seek)	-700ms (-99.3%, ~141x faster)	major architecture/path optimization
iterm session listing	302ms	303ms	+1ms (+0.3%)	effectively unchanged

Read This Before Interpreting

Workload changed: v5 includes 42 sessions (2 alive), plus correctness checks and richer enrichment paths.
Medians above are snapshot-to-snapshot: they are not from a fresh rerun in this shell.
Main structural gain is clear: JSONL path optimization removed the historical metadata bottleneck.

Conclusion

Against baseline v1, the new stack shows one major component-level improvement (JSONL version lookup) while most end-to-end operations are modestly slower in v5 (roughly +4% to +10%). This likely reflects broader enrichment scope and slightly heavier runtime paths, not a regression in the optimized JSONL path itself.

Generated: 2026-02-12 09:11:03 PST | Sources: session_bench_report_v1_baseline.html, session_bench_report_v5_20260206_120105.html